Towards Bridging Resolution in German: Data Analysis and Rule-based Experiments

Janis Pagel, Ina RoesigerNAACL 2018

TL;DR

This paper tackles the challenge of bridging resolution in German, a linguistic phenomenon where an anaphor refers to an entity not explicitly mentioned but inferentially related to a previously introduced one. The authors introduce two German corpora with bridging annotations and present a rule-based system, adapted from English, achieving modest but promising results, highlighting the importance of semantic connectivity and argument-taking features.

What Problem Does This Paper Address?

Bridging anaphora, a type of associative reference (e.g., "the documentation requirement" referring to a requirement related to the minimum wage), is a complex linguistic phenomenon. While there's been some work on bridging resolution for English, it's significantly less explored for German. This paper aims to bridge this gap by analyzing bridging phenomena in German, introducing annotated corpora, and developing and evaluating a rule-based system for identifying bridging anaphors and their antecedents in German text. This is crucial for improving natural language understanding systems, especially in tasks requiring deep semantic comprehension.

Methods

The authors adapted an existing rule-based system for bridging resolution from English to German. The system comprises three main stages:

  • Pre-processing: This involves filtering out potential anaphor candidates that are unlikely to be bridging anaphors (e.g., pronouns, indefinite expressions, proper names) and preparing gold markables.
  • Rule Application: A set of eight rules, adapted from Hou et al. (2014), are applied to identify bridging pairs. These rules leverage features like semantic connectivity (how often two words co-occur in specific grammatical patterns), argument-taking ratio (how likely a noun is to take an argument), and positional information. The authors also introduced three new rules tailored to German, focusing on country-related topics, high semantic connectivity, and political topics.
  • Post-processing: Rules are ordered by precision on a development set, and their predictions are applied to the test set.

To evaluate the effectiveness of antecedent search strategies, they also employed "oracle lists," which provide ranked suggestions of potential antecedents.

Datasets

The paper utilizes two German corpora with bridging annotations:

  • DIRNDL: A corpus of radio news transcripts, containing 655 annotated bridging pairs.
  • GRAIN: A corpus of spontaneous radio interviews, containing 274 annotated bridging pairs.

Both corpora were annotated according to the RefLex scheme, which categorizes information status. The authors note that the corefpro corpus was not available for their experiments.

Key Insights and Ideas

  • German Bridging is Challenging: The rule-based system achieved relatively low F1 scores (around 11-13%) for full bridging resolution, indicating the inherent difficulty of this task in German.
  • Semantic Connectivity is Key: Features like semantic connectivity and argument-taking ratio proved to be strong predictors for identifying bridging anaphors and their antecedents, substantiating the importance of these linguistic concepts.
  • Rule-Based Systems Benefit from Oracle Lists: While the rule-based system itself has limitations, using oracle lists (ranked suggestions of antecedents) showed potential for improving performance, especially for rules that identify correct antecedents but don't rank them highly.
  • Domain Differences Matter: The performance difference between DIRNDL (news) and GRAIN (spontaneous interviews) suggests that the nature of the discourse (scripted vs. spontaneous, topic diversity) impacts bridging resolution. GRAIN's lower semantic connectivity scores might explain its poorer performance.
  • Short, Unmodified NPs are Good Candidates: The baseline's good performance on anaphor detection suggests that many bridging anaphors are indeed short, unmodified noun phrases.

Key Results

  • Full Bridging Resolution Performance: The rule-based system achieved an F1 score of 13.6% for DIRNDL and 11.8% for GRAIN on the test set.
  • Anaphor Detection Performance: Anaphor detection scores were significantly higher, with an F1 score of 21.9% for DIRNDL and 23.5% for GRAIN on the test set.
  • Rule Performance: Individual rules showed varying levels of precision. Rules like Rule 9 (country-related) and Rule 10 (high semantic connectivity) showed promise, while others struggled to capture correct antecedents, particularly in the GRAIN corpus.
  • Coreference Information: Incorporating gold coreference information significantly improved precision and F1 scores, highlighting its importance for bridging resolution.

Limitations and Future Work

  • Low Overall Performance: The authors acknowledge the modest performance of their rule-based system, especially for full bridging resolution.
  • Rule Limitations: Many of the adapted English rules did not fire or perform well on the German data, likely due to differences in bridging definitions and corpus characteristics.
  • Semantic Connectivity Computation: The paper suggests that improving the coverage and computation of semantic connectivity is crucial for better antecedent detection, particularly for GRAIN.
  • Re-ranking Oracle Lists: The authors propose that re-ranking oracle lists could further improve performance, though this is limited by sentence distance constraints in some rules.
  • Feature Improvement: While features like semantic connectivity are promising, further improvements are needed.
  • Rule Revision: Some rules showed no improvement even with longer oracle lists, suggesting they may need revision or may not have access to correct antecedents.
  • Machine Learning Exploration: The authors mention exploring machine learning techniques but found results lower than the rule-based approach due to limited data, suggesting this as a future avenue with more data.

Summary updated

BibTeX
@inproceedings{pagel2018towards,
  author = {Janis Pagel and Ina Roesiger},
  title = {{Towards Bridging Resolution in German: Data Analysis and Rule-based Experiments}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2018},
}
Figure